Skip to main content
Meerkat Recall connects claims to the evidence that supports them. It keeps the original source, the extracted interpretation, and the current answer as separate records so that a reader can ask both what do we know? and why do we believe it? Consider a synthetic project note: “Mara joined Northstar Labs in March 2026.” That sentence can support an employment relationship and a dated event. It does not, by itself, prove that Mara still works there today.

From evidence to an answer

The diagram shows the graph-ingestion path. A source can also stop after metadata processing or document retrieval. Background work is asynchronous: accepting an ingest request does not mean its graph or wiki page is ready.

What each record means

Extraction is interpretive. The table illustrates possible records, rather than promising an exact number of rows or identical model output on every run.

Identity, evidence, and truth are different decisions

Three questions arise when another note says “Mara now works at Cedar Works”:
  1. Identity: is this the same Mara? Entity resolution uses names, identity keys, type compatibility, and, when configured, disambiguation.
  2. Vocabulary: does “works at” mean the canonical works_at predicate? Predicate registration and alias governance answer that question.
  3. Truth: did Mara change jobs, hold two roles, or appear in an ambiguous source? Truth maintenance evaluates the competing claim values and their configured semantics.
A high extraction confidence addresses what the source says. It does not settle identity or prove that the statement remains true.

Isolation and visibility

Records belong to a space. Within that space, security envelopes determine which records a principal can read or modify. Truth maintenance and knowledge views also preserve security boundaries: evidence from a more restricted envelope cannot silently become a less restricted answer. A missing search result can therefore mean that a record is unauthorized, not yet processed, excluded by a status filter, or absent. Inspect the relevant status and use the same principal and space when diagnosing it.

Choose your next page

Understand ingestion

Start with sources, revisions, and the processing depth you need.

Model a graph

Choose entity identities, predicates, and claim shapes.

Interpret an answer

Follow evidence, conflicts, validity, and staleness.

Control access

Understand spaces, clearances, subjects, and derived reads.
For a runnable starting point, use the quickstart. For exact request bodies, use the HTTP API or MCP tools.